Wie man einen KI-Sprachagenten mit Vonage Voice API und Deepgram erstellt

#==== Vonage Voice API ====
API_KEY=your_vonage_api_key
API_SECRET=your_vonage_api_secret
APP_ID=your_application_id
SERVICE_PHONE_NUMBER=your_vonage_number

#==== Deepgram Voice Agent API ====
DEEPGRAM_API_KEY=your_deepgram_api_key
DEEPGRAM_VOICE_AGENT_ENDPOINT=agent.deepgram.com/v1/agent/converse
DEEPGRAM_AGENT_SPEAK=aura-orion-en

#==== Other custom parameters ====
MAX_CALL_DURATION=300

'use strict'

require('dotenv').config();

const express = require('express');
const bodyParser = require('body-parser');
const app = express();
require('express-ws')(app);
const webSocket = require('ws');

app.use(bodyParser.json());

//---- CORS policy ----
app.use(function (req, res, next) {
  res.header("Access-Control-Allow-Origin", "*");
  res.header("Access-Control-Allow-Headers", "Origin, X-Requested-With, Content-Type, Accept");
  res.header("Access-Control-Allow-Methods", "OPTIONS,GET,POST,PUT,DELETE");
  next();
});

//---- Configuration ----
const servicePhoneNumber = process.env.SERVICE_PHONE_NUMBER;

//---- Vonage API Setup ----
const { Auth } = require('@vonage/auth');
const credentials = new Auth({
  apiKey: process.env.API_KEY,
  apiSecret: process.env.API_SECRET,
  applicationId: process.env.APP_ID,
  privateKey: './private.key'
});

const apiBaseUrl = "https://api.nexmo.com";
const options = { apiHost: apiBaseUrl };

const { Vonage } = require('@vonage/server-sdk');
const vonage = new Vonage(credentials, options);

//---- Deepgram Voice Agent Configuration ----
const dgApiKey = process.env.DEEPGRAM_API_KEY;
const dgVoiceAgentEndpoint = process.env.DEEPGRAM_VOICE_AGENT_ENDPOINT;
const dgVoiceAgentSettings = {
  "type": "Settings",
  "audio": {
    "input": { "encoding": "linear16", "sample_rate": 8000 },
    "output": { "encoding": "linear16", "sample_rate": 8000, "container": "none" }
  },
  "agent": {
    "listen": { "provider": { "type": "deepgram", "model": "nova-3" } },
    "think": {
      "provider": { "type": "anthropic", "model": "claude-sonnet-4-20250514" },
      "prompt": "You are a helpful AI assistant on a live phone call. Keep responses concise and natural for spoken conversation."
    },
    "speak": { 
      "provider": { 
        "type": "deepgram", 
        "model": process.env.DEEPGRAM_AGENT_SPEAK 
      } 
    }
  }
};

//---- Handle incoming PSTN calls ----
app.get('/answer', async (req, res) => {
  const hostName = req.hostname;
  const uuid = req.query.uuid;
  
  // For local development with ngrok, use your ngrok URL directly
  // const publicUrl = 'https://your-ngrok-url.ngrok.io';
  const wsUri = `wss://${hostName}/socket?original_uuid=${uuid}`;
  
  const nccoResponse = [
    {
      "action": "talk",
      "text": "Hello, please wait while we're connecting your call!",
      "language": "en-US",
      "style": 11
    },
    {
      "action": "connect",
      "eventType": "synchronous",
      "eventUrl": [`https://${hostName}/ws_event`],
      "from": req.query.from,
      "endpoint": [
        {
          "type": "websocket",
          "uri": wsUri,
          "content-type": "audio/l16;rate=8000",
          "headers": {}
        }
      ]
    }
  ];
  
  res.status(200).json(nccoResponse);
});

//---- Event webhook for call status ----
app.post('/event', async (req, res) => {
  res.status(200).send('Ok');
});

//---- WebSocket event handler ----
app.post('/ws_event', async (req, res) => {
  res.status(200).send('Ok');
  
  // Trigger a greeting when WebSocket is connected
  setTimeout(() => {
    if (req.body.status === 'answered') {
      vonage.voice.playTTS(req.body.uuid, {
        text: "Hello",
        language: 'en-US',
        style: 11
      })
      .then(res => console.log("Initial greeting sent"))
      .catch(err => console.error("Failed to play TTS:", err));
    }
  }, 1500);
});

//---- Start server ----
const port = process.env.PORT || 3000;
app.listen(port, () => {
  console.log(`Voice Agent application listening on port ${port}`);
  console.log(`Make sure ngrok is forwarding to this port!`);
});

//---- WebSocket Connector ----
app.ws('/socket', async (ws, req) => {
  let wsDgOpen = false; // Deepgram WebSocket ready?
  const originalUuid = req.query.original_uuid;
  
  console.log('WebSocket connected for call UUID:', originalUuid);
  
  //---- Connect to Deepgram Voice Agent ----
  console.log('Opening connection to Deepgram Voice Agent');
  const wsDg = new webSocket(`wss://${dgVoiceAgentEndpoint}`, {
    headers: { authorization: `token ${dgApiKey}` }
  });
  
  wsDg.on('error', async (event) => {
    console.log('WebSocket to Deepgram error:', event);
  });
  
  wsDg.on('open', () => {
    console.log('WebSocket to Deepgram opened');
    // Send configuration to Deepgram Voice Agent
    wsDg.send(JSON.stringify(dgVoiceAgentSettings));
    wsDgOpen = true;
  });
  
  //---- Handle messages from Deepgram ----
  wsDg.on('message', async (msg, isBinary) => {
    if (isBinary) {
      // Audio data from agent - send directly to Vonage
      ws.send(msg);
    } else {
      // Text messages (transcripts, events, etc.)
      const message = JSON.parse(msg.toString('utf8'));
      console.log(`Message from Deepgram:`, message);
      
      // Handle barge-in: clear Vonage's audio buffer when user starts speaking
      if (message.type === "UserStartedSpeaking") {
        ws.send(JSON.stringify({ action: "clear" }));
        console.log('Sent CLEAR command to Vonage');
      }
    }
  });
  
  wsDg.on('close', async () => {
    wsDgOpen = false;
    console.log("Deepgram WebSocket closed");
  });
  
  //---- Handle messages from Vonage (user audio) ----
  ws.on('message', async (msg) => {
    if (typeof msg === "string") {
      const event = JSON.parse(msg);
      console.log("Vonage event:", event.event);
      
      // The first message from Vonage is always websocket:connected
      if (event.event === "websocket:connected") {
        console.log('Vonage WebSocket established:', event['content-type']);
      }
      
      // Handle Vonage control message confirmations
      if (event.event === "websocket:cleared") {
        console.log('Vonage audio buffer cleared');
      }
    } else {
      // Binary audio data from caller - forward to Deepgram
      if (wsDgOpen) {
        wsDg.send(msg);
      }
    }
  });
  
  //---- Clean up on disconnect ----
  ws.on('close', async () => {
    wsDgOpen = false;
    wsDg.close();
    console.log("Vonage WebSocket closed");
  });
});

//---- Trigger outbound PSTN calls ----
app.get('/call', async (req, res) => {
  if (req.query.callee == null) {
    res.status(400).send('"callee" number missing as query parameter');
  } else {
    res.status(200).send('Ok');
    const hostName = req.hostname;
    
    vonage.voice.createOutboundCall({
      to: [{
        type: 'phone',
        number: req.query.callee
      }],
      from: {
        type: 'phone',
        number: servicePhoneNumber
      },
      limit: process.env.MAX_CALL_DURATION,
      answer_url: [`https://${hostName}/answer`],
      answer_method: 'GET',
      event_url: [`https://${hostName}/event`],
      event_method: 'POST'
    })
    .then(res => console.log("Outgoing PSTN call status:", res))
    .catch(err => console.error("Outgoing PSTN call error:", err));
  }
});

https://your-ngrok-url.ngrok.io/call?callee=15551234567

"think": {
  "provider": { "type": "open_ai", "model": "gpt-4o-mini" },
  "prompt": "You are a helpful AI assistant on a live phone call. Keep responses concise and natural for spoken conversation."
}

"prompt": "You are a friendly customer service representative for Acme Corp. Help users with their inquiries about our products and services. Be professional but warm."